Claude 扩展思考模式：Extended Thinking API

Extended Thinking 的基本使用

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000  # 最多允许思考 10000 token
    },
    messages=[{
        "role": "user",
        "content": "证明：√2 是无理数"
    }]
)

# 响应包含 thinking block 和 text block
for block in response.content:
    if block.type == "thinking":
        print("[思考过程]", block.thinking[:200], "...")
    elif block.type == "text":
        print("[最终答案]", block.text)

响应结构解析

# 完整响应结构
{
  "id": "msg_01Xfn...",
  "type": "message",
  "role": "assistant",
  "content": [
    {
      "type": "thinking",
      "thinking": "我需要证明 √2 是无理数...\n假设 √2 = p/q（最简分数）...\n则 2 = p²/q²...\n..."
      # 注意：thinking 内容可能不完整，是模型草稿
    },
    {
      "type": "text",
      "text": "**证明**\n\n采用反证法：假设 √2 是有理数..."
    }
  ],
  "usage": {
    "input_tokens": 30,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "output_tokens": 8420  # 包含思考 token
  }
}

budget_tokens 的影响

budget_tokens 控制模型可以用于"思考"的最大 token 数，直接影响效果和成本：

流式输出：实时显示思考过程

with client.messages.stream(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 8000},
    messages=[{"role": "user", "content": prompt}]
) as stream:
    current_block_type = None

    for event in stream:
        if hasattr(event, 'type'):
            if event.type == 'content_block_start':
                current_block_type = event.content_block.type
                if current_block_type == 'thinking':
                    print("\n🤔 [思考中...]", end="", flush=True)
                elif current_block_type == 'text':
                    print("\n\n💬 [回答]")

            elif event.type == 'content_block_delta':
                if current_block_type == 'thinking':
                    print(".", end="", flush=True)  # 思考时显示进度
                elif current_block_type == 'text':
                    print(event.delta.text, end="", flush=True)

在多轮对话中保留思考上下文

# 多轮对话中，需要将 thinking block 传回
conversation_history = []

def chat_with_thinking(user_message: str) -> str:
    conversation_history.append({
        "role": "user",
        "content": user_message
    })

    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=16000,
        thinking={"type": "enabled", "budget_tokens": 5000},
        messages=conversation_history
    )

    # 将完整响应（含 thinking block）加入历史
    conversation_history.append({
        "role": "assistant",
        "content": response.content  # 必须包含 thinking block
    })

    # 只返回 text block 给用户
    return next(b.text for b in response.content if b.type == "text")